The Bureaucratic Brain on GenAI

We are in the midst of a global creative project heralded by the loud arrival of accessible Large Language Models (LLMs) or Generative Artificial Intelligence (GenAI). In my first encounter with ChatGPT 3.5, it felt like my first exposure to the internet. There was this surreal sense of simultaneous empowerment and disempowerment. It was so obvious that this new, seemingly magical creature was promising something – but what or how exactly? Similar to learning how to be online, and how we’ve all transitioned to being permanently onlife, I was self-aware that this was going to change how I engaged with the world around me in unknown ways. Of course, that includes my policy work, and I felt an urgent need to make sense of my sudden arrival in a new world.

This piece isn’t about the specific opportunities and challenges that are in front of us, or the very important questions that face us to duly adjust governance and transparency frameworks to ensures that GenAI more beneficial than detrimental for how governments work with, and for, the publics they represent. Rather, this is my own personal reflection on some of the questions and feelings about how GenAI (and automated systems more broadly) may transform my daily tasks or my role, and what this reveals about my intrinsic impact or value as a policy agent. Acknowledging that mine is just one journey amidst this global adventure, this is a reflective piece about the core existential questions facing us and how, at a personal level, I am trying to grapple with such big, imposing waves of change.

Risk and Reward?

There are well noted concerns and risks, and Australian Governments like many have produced a number of guidance documents including this one, so I’m not going to wrestle with the risk equation, except to say that we suddenly have a massively powerful new technology that is easily available. Additionally, it’s a bit tricky to keep track of, to massively understate the concern. Ethan Mollick has noted how, considering the jagged frontier of AI, we are all a part of this process of co-learning the limitations and applications, in tandem with setting boundaries and rules. As a public servant, while cautious of applying it in my job, I feel that we all have a responsibility to be a part of learning its limitations and strengths, or we misunderstand the potential benefits (and drawbacks).

Existential challenges

More and more, I have the persistent feeling of being the custodian of soon-to-be obsolete tasks. That doesn’t mean I think my role will be replaced by Policy Bot 2025, or that we’ll all be replaced by the singularity. Maybe this is what it felt like working in a government agency somewhere across the former Soviet bloc, as glasnost and perestroika rolled through in the 80s/90s. It seems I’m just going through the motions with many daily tasks that can already be made redundant, with even a modest uptake of inhouse AI processes. But we’re not there yet, so it’s a strange limbo phase.

This leads to some big questions - existential ones:

• Where do I have actual impact and add value in my role?

• Do redundant tasks need to be automated, or are some of the automatable ones ripe for deletion entirely?

• Where is my attention most of the day – how does the work actually get done?

• What is government even supposed to be about, anyway?

The notion of time allocation cuts deep, both in terms of how I am or can be more effective in my daily efforts, but also where I put my time with ad hoc GenAI testing. It’s not enough to uncritically recreate automated versions of our current processes, we need to contend with how the roles themselves can or should (and should not) be entirely redefined. What individual agency and responsibility do we have to shape that process for the better?

In my view, we can think about this transition at three levels:

• Individual: how can I improve my daily tasks, including redefining or jettisoning them where appropriate, as well as articulate where value is (or is not) added.
• Team: what opportunities are there to explore and collaborate together, particularly in sharing and testing GenAI lessons, across a work unit.
• Organisation: where/how can organisations proactively share and incentivise prototyping environments for safe testing; where/how can they lead the critical strategic work of reimaging the entire operating system, from the vantage point of evaluating (potential) structural and functional shifts.

The great age of question making

If you’ve been hovering around this technology, it feels like philosophy has become democratised. LLMs open up a raft of why, how, why not questions once you witness some of the things it can do (and its limitations!!) – including reflecting on our strengths as public servants.

In policy, traditionally we’ve been looked to for information and answers. However, I think it’s clear between LLMs and the internet, that our real power is context-awareness and question asking, rather than simply holding information. In the current digital age, my daily job involves transacting huge volumes of information with other government actors that have different priorities and contexts. The sheer magnitude of this information means that even within your own specialised domain, you can only ever possess a small parcel of all information. I think it’s time to relinquish that delusion that this can actually be overcome. Instead, I’d rather put my energy into framing stickier, higher order questions that directly serve my own context needs, which necessitates an acceptance that knowing everything is futile (but also unnecessary). It’s a relief to give up the struggle of drowning under information overload, and focus more on being a filter rather than a human database.

As “every data collection method is constrained and every dataset is filtered”, if our role becomes more about retrieving information, and sense-making informed by skilled questioning, we have an increased responsibility to engage with the inherent limitations of data and information. Our value here is both in investing more attention towards examining the foundational questions and assumptions we apply, and also being more rigorous in interrogating any emergent answers. This involves both an appraisal, again, of what our core purpose or relevance as bureaucrats is, as well as a need to enhance basic data literacy across the public sector.

Exploring together

While these are just my reflections on a sporadic immersion over the last 18 months, I’m aware that there is a need to become more strategic - and more methodical - about how to approach this adaptation. GenAI’s arrival challenges us to think more about data quality, data compression and lossiness.

The bureaucracy itself is a sophisticated stack of algorithms, they just happen to be mostly humans (currently). Policy action only ever happens when the whole complexity of the world is compressed into pragmatic, if restricted, models. So in my view, a big part of learning how to cope with applying and living with GenAI is reflecting on how bureaucracies and democracy have accrued various assurance, trust and implementation frameworks over time, and how these enable government action, while better understanding the unavoidable trade-offs that come with compressing so much information.

(Image credit: Unsplash)